1 //=====-- AMDGPUSubtarget.h - Define Subtarget for AMDGPU ------*- C++ -*-====// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //==-----------------------------------------------------------------------===// 8 // 9 /// \file 10 /// AMDGPU specific subclass of TargetSubtarget. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #ifndef LLVM_LIB_TARGET_AMDGPU_AMDGPUSUBTARGET_H 15 #define LLVM_LIB_TARGET_AMDGPU_AMDGPUSUBTARGET_H 16 17 #include "AMDGPU.h" 18 #include "AMDGPUCallLowering.h" 19 #include "R600FrameLowering.h" 20 #include "R600ISelLowering.h" 21 #include "R600InstrInfo.h" 22 #include "SIFrameLowering.h" 23 #include "SIISelLowering.h" 24 #include "SIInstrInfo.h" 25 #include "Utils/AMDGPUBaseInfo.h" 26 #include "llvm/ADT/Triple.h" 27 #include "llvm/CodeGen/GlobalISel/InstructionSelector.h" 28 #include "llvm/CodeGen/GlobalISel/LegalizerInfo.h" 29 #include "llvm/CodeGen/GlobalISel/RegisterBankInfo.h" 30 #include "llvm/CodeGen/MachineFunction.h" 31 #include "llvm/CodeGen/SelectionDAGTargetInfo.h" 32 #include "llvm/MC/MCInstrItineraries.h" 33 #include "llvm/Support/MathExtras.h" 34 #include <cassert> 35 #include <cstdint> 36 #include <memory> 37 #include <utility> 38 39 #define GET_SUBTARGETINFO_HEADER 40 #include "AMDGPUGenSubtargetInfo.inc" 41 #define GET_SUBTARGETINFO_HEADER 42 #include "R600GenSubtargetInfo.inc" 43 44 namespace llvm { 45 46 class StringRef; 47 48 class AMDGPUSubtarget { 49 public: 50 enum Generation { 51 R600 = 0, 52 R700 = 1, 53 EVERGREEN = 2, 54 NORTHERN_ISLANDS = 3, 55 SOUTHERN_ISLANDS = 4, 56 SEA_ISLANDS = 5, 57 VOLCANIC_ISLANDS = 6, 58 GFX9 = 7, 59 GFX10 = 8 60 }; 61 62 private: 63 Triple TargetTriple; 64 65 protected: 66 bool Has16BitInsts; 67 bool HasMadMixInsts; 68 bool FP32Denormals; 69 bool FPExceptions; 70 bool HasSDWA; 71 bool HasVOP3PInsts; 72 bool HasMulI24; 73 bool HasMulU24; 74 bool HasInv2PiInlineImm; 75 bool HasFminFmaxLegacy; 76 bool EnablePromoteAlloca; 77 bool HasTrigReducedRange; 78 unsigned MaxWavesPerEU; 79 int LocalMemorySize; 80 unsigned WavefrontSize; 81 82 public: 83 AMDGPUSubtarget(const Triple &TT); 84 85 static const AMDGPUSubtarget &get(const MachineFunction &MF); 86 static const AMDGPUSubtarget &get(const TargetMachine &TM, 87 const Function &F); 88 89 /// \returns Default range flat work group size for a calling convention. 90 std::pair<unsigned, unsigned> getDefaultFlatWorkGroupSize(CallingConv::ID CC) const; 91 92 /// \returns Subtarget's default pair of minimum/maximum flat work group sizes 93 /// for function \p F, or minimum/maximum flat work group sizes explicitly 94 /// requested using "amdgpu-flat-work-group-size" attribute attached to 95 /// function \p F. 96 /// 97 /// \returns Subtarget's default values if explicitly requested values cannot 98 /// be converted to integer, or violate subtarget's specifications. 99 std::pair<unsigned, unsigned> getFlatWorkGroupSizes(const Function &F) const; 100 101 /// \returns Subtarget's default pair of minimum/maximum number of waves per 102 /// execution unit for function \p F, or minimum/maximum number of waves per 103 /// execution unit explicitly requested using "amdgpu-waves-per-eu" attribute 104 /// attached to function \p F. 105 /// 106 /// \returns Subtarget's default values if explicitly requested values cannot 107 /// be converted to integer, violate subtarget's specifications, or are not 108 /// compatible with minimum/maximum number of waves limited by flat work group 109 /// size, register usage, and/or lds usage. 110 std::pair<unsigned, unsigned> getWavesPerEU(const Function &F) const; 111 112 /// Return the amount of LDS that can be used that will not restrict the 113 /// occupancy lower than WaveCount. 114 unsigned getMaxLocalMemSizeWithWaveCount(unsigned WaveCount, 115 const Function &) const; 116 117 /// Inverse of getMaxLocalMemWithWaveCount. Return the maximum wavecount if 118 /// the given LDS memory size is the only constraint. 119 unsigned getOccupancyWithLocalMemSize(uint32_t Bytes, const Function &) const; 120 121 unsigned getOccupancyWithLocalMemSize(const MachineFunction &MF) const; 122 123 bool isAmdHsaOS() const { 124 return TargetTriple.getOS() == Triple::AMDHSA; 125 } 126 127 bool isAmdPalOS() const { 128 return TargetTriple.getOS() == Triple::AMDPAL; 129 } 130 131 bool isMesa3DOS() const { 132 return TargetTriple.getOS() == Triple::Mesa3D; 133 } 134 135 bool isMesaKernel(const Function &F) const { 136 return isMesa3DOS() && !AMDGPU::isShader(F.getCallingConv()); 137 } 138 139 bool isAmdHsaOrMesa(const Function &F) const { 140 return isAmdHsaOS() || isMesaKernel(F); 141 } 142 143 bool has16BitInsts() const { 144 return Has16BitInsts; 145 } 146 147 bool hasMadMixInsts() const { 148 return HasMadMixInsts; 149 } 150 151 bool hasFP32Denormals(const Function &F) const { 152 // FIXME: This should not be a property of the subtarget. This should be a 153 // property with a default set by the calling convention which can be 154 // overridden by attributes. For now, use the subtarget feature as a 155 // placeholder attribute. The function arguments only purpose is to 156 // discourage use without a function context until this is removed. 157 return FP32Denormals; 158 } 159 160 bool hasFPExceptions() const { 161 return FPExceptions; 162 } 163 164 bool hasSDWA() const { 165 return HasSDWA; 166 } 167 168 bool hasVOP3PInsts() const { 169 return HasVOP3PInsts; 170 } 171 172 bool hasMulI24() const { 173 return HasMulI24; 174 } 175 176 bool hasMulU24() const { 177 return HasMulU24; 178 } 179 180 bool hasInv2PiInlineImm() const { 181 return HasInv2PiInlineImm; 182 } 183 184 bool hasFminFmaxLegacy() const { 185 return HasFminFmaxLegacy; 186 } 187 188 bool hasTrigReducedRange() const { 189 return HasTrigReducedRange; 190 } 191 192 bool isPromoteAllocaEnabled() const { 193 return EnablePromoteAlloca; 194 } 195 196 unsigned getWavefrontSize() const { 197 return WavefrontSize; 198 } 199 200 int getLocalMemorySize() const { 201 return LocalMemorySize; 202 } 203 204 Align getAlignmentForImplicitArgPtr() const { 205 return isAmdHsaOS() ? Align(8) : Align(4); 206 } 207 208 /// Returns the offset in bytes from the start of the input buffer 209 /// of the first explicit kernel argument. 210 unsigned getExplicitKernelArgOffset(const Function &F) const { 211 return isAmdHsaOrMesa(F) ? 0 : 36; 212 } 213 214 /// \returns Maximum number of work groups per compute unit supported by the 215 /// subtarget and limited by given \p FlatWorkGroupSize. 216 virtual unsigned getMaxWorkGroupsPerCU(unsigned FlatWorkGroupSize) const = 0; 217 218 /// \returns Minimum flat work group size supported by the subtarget. 219 virtual unsigned getMinFlatWorkGroupSize() const = 0; 220 221 /// \returns Maximum flat work group size supported by the subtarget. 222 virtual unsigned getMaxFlatWorkGroupSize() const = 0; 223 224 /// \returns Maximum number of waves per execution unit supported by the 225 /// subtarget and limited by given \p FlatWorkGroupSize. 226 virtual unsigned getMaxWavesPerEU(unsigned FlatWorkGroupSize) const = 0; 227 228 /// \returns Minimum number of waves per execution unit supported by the 229 /// subtarget. 230 virtual unsigned getMinWavesPerEU() const = 0; 231 232 /// \returns Maximum number of waves per execution unit supported by the 233 /// subtarget without any kind of limitation. 234 unsigned getMaxWavesPerEU() const { return MaxWavesPerEU; } 235 236 /// Creates value range metadata on an workitemid.* inrinsic call or load. 237 bool makeLIDRangeMetadata(Instruction *I) const; 238 239 /// \returns Number of bytes of arguments that are passed to a shader or 240 /// kernel in addition to the explicit ones declared for the function. 241 unsigned getImplicitArgNumBytes(const Function &F) const { 242 if (isMesaKernel(F)) 243 return 16; 244 return AMDGPU::getIntegerAttribute(F, "amdgpu-implicitarg-num-bytes", 0); 245 } 246 uint64_t getExplicitKernArgSize(const Function &F, Align &MaxAlign) const; 247 unsigned getKernArgSegmentSize(const Function &F, Align &MaxAlign) const; 248 249 virtual ~AMDGPUSubtarget() {} 250 }; 251 252 class GCNSubtarget : public AMDGPUGenSubtargetInfo, 253 public AMDGPUSubtarget { 254 255 using AMDGPUSubtarget::getMaxWavesPerEU; 256 257 public: 258 enum TrapHandlerAbi { 259 TrapHandlerAbiNone = 0, 260 TrapHandlerAbiHsa = 1 261 }; 262 263 enum TrapID { 264 TrapIDHardwareReserved = 0, 265 TrapIDHSADebugTrap = 1, 266 TrapIDLLVMTrap = 2, 267 TrapIDLLVMDebugTrap = 3, 268 TrapIDDebugBreakpoint = 7, 269 TrapIDDebugReserved8 = 8, 270 TrapIDDebugReservedFE = 0xfe, 271 TrapIDDebugReservedFF = 0xff 272 }; 273 274 enum TrapRegValues { 275 LLVMTrapHandlerRegValue = 1 276 }; 277 278 private: 279 /// GlobalISel related APIs. 280 std::unique_ptr<AMDGPUCallLowering> CallLoweringInfo; 281 std::unique_ptr<InstructionSelector> InstSelector; 282 std::unique_ptr<LegalizerInfo> Legalizer; 283 std::unique_ptr<RegisterBankInfo> RegBankInfo; 284 285 protected: 286 // Basic subtarget description. 287 Triple TargetTriple; 288 unsigned Gen; 289 InstrItineraryData InstrItins; 290 int LDSBankCount; 291 unsigned MaxPrivateElementSize; 292 293 // Possibly statically set by tablegen, but may want to be overridden. 294 bool FastFMAF32; 295 bool HalfRate64Ops; 296 297 // Dynamially set bits that enable features. 298 bool FP64FP16Denormals; 299 bool FlatForGlobal; 300 bool AutoWaitcntBeforeBarrier; 301 bool CodeObjectV3; 302 bool UnalignedScratchAccess; 303 bool UnalignedBufferAccess; 304 bool HasApertureRegs; 305 bool EnableXNACK; 306 bool DoesNotSupportXNACK; 307 bool EnableCuMode; 308 bool TrapHandler; 309 310 // Used as options. 311 bool EnableLoadStoreOpt; 312 bool EnableUnsafeDSOffsetFolding; 313 bool EnableSIScheduler; 314 bool EnableDS128; 315 bool EnablePRTStrictNull; 316 bool DumpCode; 317 318 // Subtarget statically properties set by tablegen 319 bool FP64; 320 bool FMA; 321 bool MIMG_R128; 322 bool IsGCN; 323 bool GCN3Encoding; 324 bool CIInsts; 325 bool GFX8Insts; 326 bool GFX9Insts; 327 bool GFX10Insts; 328 bool GFX7GFX8GFX9Insts; 329 bool SGPRInitBug; 330 bool HasSMemRealTime; 331 bool HasIntClamp; 332 bool HasFmaMixInsts; 333 bool HasMovrel; 334 bool HasVGPRIndexMode; 335 bool HasScalarStores; 336 bool HasScalarAtomics; 337 bool HasSDWAOmod; 338 bool HasSDWAScalar; 339 bool HasSDWASdst; 340 bool HasSDWAMac; 341 bool HasSDWAOutModsVOPC; 342 bool HasDPP; 343 bool HasDPP8; 344 bool HasR128A16; 345 bool HasNSAEncoding; 346 bool HasDLInsts; 347 bool HasDot1Insts; 348 bool HasDot2Insts; 349 bool HasDot3Insts; 350 bool HasDot4Insts; 351 bool HasDot5Insts; 352 bool HasDot6Insts; 353 bool HasMAIInsts; 354 bool HasPkFmacF16Inst; 355 bool HasAtomicFaddInsts; 356 bool EnableSRAMECC; 357 bool DoesNotSupportSRAMECC; 358 bool HasNoSdstCMPX; 359 bool HasVscnt; 360 bool HasRegisterBanking; 361 bool HasVOP3Literal; 362 bool HasNoDataDepHazard; 363 bool FlatAddressSpace; 364 bool FlatInstOffsets; 365 bool FlatGlobalInsts; 366 bool FlatScratchInsts; 367 bool ScalarFlatScratchInsts; 368 bool AddNoCarryInsts; 369 bool HasUnpackedD16VMem; 370 bool R600ALUInst; 371 bool CaymanISA; 372 bool CFALUBug; 373 bool LDSMisalignedBug; 374 bool HasMFMAInlineLiteralBug; 375 bool HasVertexCache; 376 short TexVTXClauseSize; 377 bool ScalarizeGlobal; 378 379 bool HasVcmpxPermlaneHazard; 380 bool HasVMEMtoScalarWriteHazard; 381 bool HasSMEMtoVectorWriteHazard; 382 bool HasInstFwdPrefetchBug; 383 bool HasVcmpxExecWARHazard; 384 bool HasLdsBranchVmemWARHazard; 385 bool HasNSAtoVMEMBug; 386 bool HasOffset3fBug; 387 bool HasFlatSegmentOffsetBug; 388 389 // Dummy feature to use for assembler in tablegen. 390 bool FeatureDisable; 391 392 SelectionDAGTargetInfo TSInfo; 393 private: 394 SIInstrInfo InstrInfo; 395 SITargetLowering TLInfo; 396 SIFrameLowering FrameLowering; 397 398 // See COMPUTE_TMPRING_SIZE.WAVESIZE, 13-bit field in units of 256-dword. 399 static const unsigned MaxWaveScratchSize = (256 * 4) * ((1 << 13) - 1); 400 401 public: 402 GCNSubtarget(const Triple &TT, StringRef GPU, StringRef FS, 403 const GCNTargetMachine &TM); 404 ~GCNSubtarget() override; 405 406 GCNSubtarget &initializeSubtargetDependencies(const Triple &TT, 407 StringRef GPU, StringRef FS); 408 409 const SIInstrInfo *getInstrInfo() const override { 410 return &InstrInfo; 411 } 412 413 const SIFrameLowering *getFrameLowering() const override { 414 return &FrameLowering; 415 } 416 417 const SITargetLowering *getTargetLowering() const override { 418 return &TLInfo; 419 } 420 421 const SIRegisterInfo *getRegisterInfo() const override { 422 return &InstrInfo.getRegisterInfo(); 423 } 424 425 const CallLowering *getCallLowering() const override { 426 return CallLoweringInfo.get(); 427 } 428 429 InstructionSelector *getInstructionSelector() const override { 430 return InstSelector.get(); 431 } 432 433 const LegalizerInfo *getLegalizerInfo() const override { 434 return Legalizer.get(); 435 } 436 437 const RegisterBankInfo *getRegBankInfo() const override { 438 return RegBankInfo.get(); 439 } 440 441 // Nothing implemented, just prevent crashes on use. 442 const SelectionDAGTargetInfo *getSelectionDAGInfo() const override { 443 return &TSInfo; 444 } 445 446 const InstrItineraryData *getInstrItineraryData() const override { 447 return &InstrItins; 448 } 449 450 void ParseSubtargetFeatures(StringRef CPU, StringRef FS); 451 452 Generation getGeneration() const { 453 return (Generation)Gen; 454 } 455 456 unsigned getWavefrontSizeLog2() const { 457 return Log2_32(WavefrontSize); 458 } 459 460 /// Return the number of high bits known to be zero fror a frame index. 461 unsigned getKnownHighZeroBitsForFrameIndex() const { 462 return countLeadingZeros(MaxWaveScratchSize) + getWavefrontSizeLog2(); 463 } 464 465 int getLDSBankCount() const { 466 return LDSBankCount; 467 } 468 469 unsigned getMaxPrivateElementSize() const { 470 return MaxPrivateElementSize; 471 } 472 473 unsigned getConstantBusLimit(unsigned Opcode) const; 474 475 bool hasIntClamp() const { 476 return HasIntClamp; 477 } 478 479 bool hasFP64() const { 480 return FP64; 481 } 482 483 bool hasMIMG_R128() const { 484 return MIMG_R128; 485 } 486 487 bool hasHWFP64() const { 488 return FP64; 489 } 490 491 bool hasFastFMAF32() const { 492 return FastFMAF32; 493 } 494 495 bool hasHalfRate64Ops() const { 496 return HalfRate64Ops; 497 } 498 499 bool hasAddr64() const { 500 return (getGeneration() < AMDGPUSubtarget::VOLCANIC_ISLANDS); 501 } 502 503 // Return true if the target only has the reverse operand versions of VALU 504 // shift instructions (e.g. v_lshrrev_b32, and no v_lshr_b32). 505 bool hasOnlyRevVALUShifts() const { 506 return getGeneration() >= VOLCANIC_ISLANDS; 507 } 508 509 bool hasBFE() const { 510 return true; 511 } 512 513 bool hasBFI() const { 514 return true; 515 } 516 517 bool hasBFM() const { 518 return hasBFE(); 519 } 520 521 bool hasBCNT(unsigned Size) const { 522 return true; 523 } 524 525 bool hasFFBL() const { 526 return true; 527 } 528 529 bool hasFFBH() const { 530 return true; 531 } 532 533 bool hasMed3_16() const { 534 return getGeneration() >= AMDGPUSubtarget::GFX9; 535 } 536 537 bool hasMin3Max3_16() const { 538 return getGeneration() >= AMDGPUSubtarget::GFX9; 539 } 540 541 bool hasFmaMixInsts() const { 542 return HasFmaMixInsts; 543 } 544 545 bool hasCARRY() const { 546 return true; 547 } 548 549 bool hasFMA() const { 550 return FMA; 551 } 552 553 bool hasSwap() const { 554 return GFX9Insts; 555 } 556 557 bool hasScalarPackInsts() const { 558 return GFX9Insts; 559 } 560 561 bool hasScalarMulHiInsts() const { 562 return GFX9Insts; 563 } 564 565 TrapHandlerAbi getTrapHandlerAbi() const { 566 return isAmdHsaOS() ? TrapHandlerAbiHsa : TrapHandlerAbiNone; 567 } 568 569 /// True if the offset field of DS instructions works as expected. On SI, the 570 /// offset uses a 16-bit adder and does not always wrap properly. 571 bool hasUsableDSOffset() const { 572 return getGeneration() >= SEA_ISLANDS; 573 } 574 575 bool unsafeDSOffsetFoldingEnabled() const { 576 return EnableUnsafeDSOffsetFolding; 577 } 578 579 /// Condition output from div_scale is usable. 580 bool hasUsableDivScaleConditionOutput() const { 581 return getGeneration() != SOUTHERN_ISLANDS; 582 } 583 584 /// Extra wait hazard is needed in some cases before 585 /// s_cbranch_vccnz/s_cbranch_vccz. 586 bool hasReadVCCZBug() const { 587 return getGeneration() <= SEA_ISLANDS; 588 } 589 590 /// Writes to VCC_LO/VCC_HI update the VCCZ flag. 591 bool partialVCCWritesUpdateVCCZ() const { 592 return getGeneration() >= GFX10; 593 } 594 595 /// A read of an SGPR by SMRD instruction requires 4 wait states when the SGPR 596 /// was written by a VALU instruction. 597 bool hasSMRDReadVALUDefHazard() const { 598 return getGeneration() == SOUTHERN_ISLANDS; 599 } 600 601 /// A read of an SGPR by a VMEM instruction requires 5 wait states when the 602 /// SGPR was written by a VALU Instruction. 603 bool hasVMEMReadSGPRVALUDefHazard() const { 604 return getGeneration() >= VOLCANIC_ISLANDS; 605 } 606 607 bool hasRFEHazards() const { 608 return getGeneration() >= VOLCANIC_ISLANDS; 609 } 610 611 /// Number of hazard wait states for s_setreg_b32/s_setreg_imm32_b32. 612 unsigned getSetRegWaitStates() const { 613 return getGeneration() <= SEA_ISLANDS ? 1 : 2; 614 } 615 616 bool dumpCode() const { 617 return DumpCode; 618 } 619 620 /// Return the amount of LDS that can be used that will not restrict the 621 /// occupancy lower than WaveCount. 622 unsigned getMaxLocalMemSizeWithWaveCount(unsigned WaveCount, 623 const Function &) const; 624 625 /// Alias for hasFP64FP16Denormals 626 bool hasFP16Denormals(const Function &F) const { 627 return FP64FP16Denormals; 628 } 629 630 /// Alias for hasFP64FP16Denormals 631 bool hasFP64Denormals(const Function &F) const { 632 return FP64FP16Denormals; 633 } 634 635 bool hasFP64FP16Denormals(const Function &F) const { 636 return FP64FP16Denormals; 637 } 638 639 bool supportsMinMaxDenormModes() const { 640 return getGeneration() >= AMDGPUSubtarget::GFX9; 641 } 642 643 /// \returns If target supports S_DENORM_MODE. 644 bool hasDenormModeInst() const { 645 return getGeneration() >= AMDGPUSubtarget::GFX10; 646 } 647 648 bool useFlatForGlobal() const { 649 return FlatForGlobal; 650 } 651 652 /// \returns If target supports ds_read/write_b128 and user enables generation 653 /// of ds_read/write_b128. 654 bool useDS128() const { 655 return CIInsts && EnableDS128; 656 } 657 658 /// Have v_trunc_f64, v_ceil_f64, v_rndne_f64 659 bool haveRoundOpsF64() const { 660 return CIInsts; 661 } 662 663 /// \returns If MUBUF instructions always perform range checking, even for 664 /// buffer resources used for private memory access. 665 bool privateMemoryResourceIsRangeChecked() const { 666 return getGeneration() < AMDGPUSubtarget::GFX9; 667 } 668 669 /// \returns If target requires PRT Struct NULL support (zero result registers 670 /// for sparse texture support). 671 bool usePRTStrictNull() const { 672 return EnablePRTStrictNull; 673 } 674 675 bool hasAutoWaitcntBeforeBarrier() const { 676 return AutoWaitcntBeforeBarrier; 677 } 678 679 bool hasCodeObjectV3() const { 680 // FIXME: Need to add code object v3 support for mesa and pal. 681 return isAmdHsaOS() ? CodeObjectV3 : false; 682 } 683 684 bool hasUnalignedBufferAccess() const { 685 return UnalignedBufferAccess; 686 } 687 688 bool hasUnalignedScratchAccess() const { 689 return UnalignedScratchAccess; 690 } 691 692 bool hasApertureRegs() const { 693 return HasApertureRegs; 694 } 695 696 bool isTrapHandlerEnabled() const { 697 return TrapHandler; 698 } 699 700 bool isXNACKEnabled() const { 701 return EnableXNACK; 702 } 703 704 bool isCuModeEnabled() const { 705 return EnableCuMode; 706 } 707 708 bool hasFlatAddressSpace() const { 709 return FlatAddressSpace; 710 } 711 712 bool hasFlatScrRegister() const { 713 return hasFlatAddressSpace(); 714 } 715 716 bool hasFlatInstOffsets() const { 717 return FlatInstOffsets; 718 } 719 720 bool hasFlatGlobalInsts() const { 721 return FlatGlobalInsts; 722 } 723 724 bool hasFlatScratchInsts() const { 725 return FlatScratchInsts; 726 } 727 728 bool hasScalarFlatScratchInsts() const { 729 return ScalarFlatScratchInsts; 730 } 731 732 bool hasMultiDwordFlatScratchAddressing() const { 733 return getGeneration() >= GFX9; 734 } 735 736 bool hasFlatSegmentOffsetBug() const { 737 return HasFlatSegmentOffsetBug; 738 } 739 740 bool hasFlatLgkmVMemCountInOrder() const { 741 return getGeneration() > GFX9; 742 } 743 744 bool hasD16LoadStore() const { 745 return getGeneration() >= GFX9; 746 } 747 748 bool d16PreservesUnusedBits() const { 749 return hasD16LoadStore() && !isSRAMECCEnabled(); 750 } 751 752 bool hasD16Images() const { 753 return getGeneration() >= VOLCANIC_ISLANDS; 754 } 755 756 /// Return if most LDS instructions have an m0 use that require m0 to be 757 /// iniitalized. 758 bool ldsRequiresM0Init() const { 759 return getGeneration() < GFX9; 760 } 761 762 // True if the hardware rewinds and replays GWS operations if a wave is 763 // preempted. 764 // 765 // If this is false, a GWS operation requires testing if a nack set the 766 // MEM_VIOL bit, and repeating if so. 767 bool hasGWSAutoReplay() const { 768 return getGeneration() >= GFX9; 769 } 770 771 /// \returns if target has ds_gws_sema_release_all instruction. 772 bool hasGWSSemaReleaseAll() const { 773 return CIInsts; 774 } 775 776 bool hasAddNoCarry() const { 777 return AddNoCarryInsts; 778 } 779 780 bool hasUnpackedD16VMem() const { 781 return HasUnpackedD16VMem; 782 } 783 784 // Covers VS/PS/CS graphics shaders 785 bool isMesaGfxShader(const Function &F) const { 786 return isMesa3DOS() && AMDGPU::isShader(F.getCallingConv()); 787 } 788 789 bool hasMad64_32() const { 790 return getGeneration() >= SEA_ISLANDS; 791 } 792 793 bool hasSDWAOmod() const { 794 return HasSDWAOmod; 795 } 796 797 bool hasSDWAScalar() const { 798 return HasSDWAScalar; 799 } 800 801 bool hasSDWASdst() const { 802 return HasSDWASdst; 803 } 804 805 bool hasSDWAMac() const { 806 return HasSDWAMac; 807 } 808 809 bool hasSDWAOutModsVOPC() const { 810 return HasSDWAOutModsVOPC; 811 } 812 813 bool hasDLInsts() const { 814 return HasDLInsts; 815 } 816 817 bool hasDot1Insts() const { 818 return HasDot1Insts; 819 } 820 821 bool hasDot2Insts() const { 822 return HasDot2Insts; 823 } 824 825 bool hasDot3Insts() const { 826 return HasDot3Insts; 827 } 828 829 bool hasDot4Insts() const { 830 return HasDot4Insts; 831 } 832 833 bool hasDot5Insts() const { 834 return HasDot5Insts; 835 } 836 837 bool hasDot6Insts() const { 838 return HasDot6Insts; 839 } 840 841 bool hasMAIInsts() const { 842 return HasMAIInsts; 843 } 844 845 bool hasPkFmacF16Inst() const { 846 return HasPkFmacF16Inst; 847 } 848 849 bool hasAtomicFaddInsts() const { 850 return HasAtomicFaddInsts; 851 } 852 853 bool isSRAMECCEnabled() const { 854 return EnableSRAMECC; 855 } 856 857 bool hasNoSdstCMPX() const { 858 return HasNoSdstCMPX; 859 } 860 861 bool hasVscnt() const { 862 return HasVscnt; 863 } 864 865 bool hasRegisterBanking() const { 866 return HasRegisterBanking; 867 } 868 869 bool hasVOP3Literal() const { 870 return HasVOP3Literal; 871 } 872 873 bool hasNoDataDepHazard() const { 874 return HasNoDataDepHazard; 875 } 876 877 bool vmemWriteNeedsExpWaitcnt() const { 878 return getGeneration() < SEA_ISLANDS; 879 } 880 881 // Scratch is allocated in 256 dword per wave blocks for the entire 882 // wavefront. When viewed from the perspecive of an arbitrary workitem, this 883 // is 4-byte aligned. 884 // 885 // Only 4-byte alignment is really needed to access anything. Transformations 886 // on the pointer value itself may rely on the alignment / known low bits of 887 // the pointer. Set this to something above the minimum to avoid needing 888 // dynamic realignment in common cases. 889 Align getStackAlignment() const { return Align(16); } 890 891 bool enableMachineScheduler() const override { 892 return true; 893 } 894 895 bool enableSubRegLiveness() const override { 896 return true; 897 } 898 899 void setScalarizeGlobalBehavior(bool b) { ScalarizeGlobal = b; } 900 bool getScalarizeGlobalBehavior() const { return ScalarizeGlobal; } 901 902 /// \returns Number of execution units per compute unit supported by the 903 /// subtarget. 904 unsigned getEUsPerCU() const { 905 return AMDGPU::IsaInfo::getEUsPerCU(this); 906 } 907 908 /// \returns Maximum number of waves per compute unit supported by the 909 /// subtarget without any kind of limitation. 910 unsigned getMaxWavesPerCU() const { 911 return AMDGPU::IsaInfo::getMaxWavesPerCU(this); 912 } 913 914 /// \returns Maximum number of waves per compute unit supported by the 915 /// subtarget and limited by given \p FlatWorkGroupSize. 916 unsigned getMaxWavesPerCU(unsigned FlatWorkGroupSize) const { 917 return AMDGPU::IsaInfo::getMaxWavesPerCU(this, FlatWorkGroupSize); 918 } 919 920 /// \returns Number of waves per work group supported by the subtarget and 921 /// limited by given \p FlatWorkGroupSize. 922 unsigned getWavesPerWorkGroup(unsigned FlatWorkGroupSize) const { 923 return AMDGPU::IsaInfo::getWavesPerWorkGroup(this, FlatWorkGroupSize); 924 } 925 926 // static wrappers 927 static bool hasHalfRate64Ops(const TargetSubtargetInfo &STI); 928 929 // XXX - Why is this here if it isn't in the default pass set? 930 bool enableEarlyIfConversion() const override { 931 return true; 932 } 933 934 void overrideSchedPolicy(MachineSchedPolicy &Policy, 935 unsigned NumRegionInstrs) const override; 936 937 unsigned getMaxNumUserSGPRs() const { 938 return 16; 939 } 940 941 bool hasSMemRealTime() const { 942 return HasSMemRealTime; 943 } 944 945 bool hasMovrel() const { 946 return HasMovrel; 947 } 948 949 bool hasVGPRIndexMode() const { 950 return HasVGPRIndexMode; 951 } 952 953 bool useVGPRIndexMode() const; 954 955 bool hasScalarCompareEq64() const { 956 return getGeneration() >= VOLCANIC_ISLANDS; 957 } 958 959 bool hasScalarStores() const { 960 return HasScalarStores; 961 } 962 963 bool hasScalarAtomics() const { 964 return HasScalarAtomics; 965 } 966 967 bool hasLDSFPAtomics() const { 968 return GFX8Insts; 969 } 970 971 bool hasDPP() const { 972 return HasDPP; 973 } 974 975 bool hasDPPBroadcasts() const { 976 return HasDPP && getGeneration() < GFX10; 977 } 978 979 bool hasDPPWavefrontShifts() const { 980 return HasDPP && getGeneration() < GFX10; 981 } 982 983 bool hasDPP8() const { 984 return HasDPP8; 985 } 986 987 bool hasR128A16() const { 988 return HasR128A16; 989 } 990 991 bool hasOffset3fBug() const { 992 return HasOffset3fBug; 993 } 994 995 bool hasNSAEncoding() const { 996 return HasNSAEncoding; 997 } 998 999 bool hasMadF16() const; 1000 1001 bool enableSIScheduler() const { 1002 return EnableSIScheduler; 1003 } 1004 1005 bool loadStoreOptEnabled() const { 1006 return EnableLoadStoreOpt; 1007 } 1008 1009 bool hasSGPRInitBug() const { 1010 return SGPRInitBug; 1011 } 1012 1013 bool hasMFMAInlineLiteralBug() const { 1014 return HasMFMAInlineLiteralBug; 1015 } 1016 1017 bool has12DWordStoreHazard() const { 1018 return getGeneration() != AMDGPUSubtarget::SOUTHERN_ISLANDS; 1019 } 1020 1021 // \returns true if the subtarget supports DWORDX3 load/store instructions. 1022 bool hasDwordx3LoadStores() const { 1023 return CIInsts; 1024 } 1025 1026 bool hasSMovFedHazard() const { 1027 return getGeneration() == AMDGPUSubtarget::GFX9; 1028 } 1029 1030 bool hasReadM0MovRelInterpHazard() const { 1031 return getGeneration() == AMDGPUSubtarget::GFX9; 1032 } 1033 1034 bool hasReadM0SendMsgHazard() const { 1035 return getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS && 1036 getGeneration() <= AMDGPUSubtarget::GFX9; 1037 } 1038 1039 bool hasVcmpxPermlaneHazard() const { 1040 return HasVcmpxPermlaneHazard; 1041 } 1042 1043 bool hasVMEMtoScalarWriteHazard() const { 1044 return HasVMEMtoScalarWriteHazard; 1045 } 1046 1047 bool hasSMEMtoVectorWriteHazard() const { 1048 return HasSMEMtoVectorWriteHazard; 1049 } 1050 1051 bool hasLDSMisalignedBug() const { 1052 return LDSMisalignedBug && !EnableCuMode; 1053 } 1054 1055 bool hasInstFwdPrefetchBug() const { 1056 return HasInstFwdPrefetchBug; 1057 } 1058 1059 bool hasVcmpxExecWARHazard() const { 1060 return HasVcmpxExecWARHazard; 1061 } 1062 1063 bool hasLdsBranchVmemWARHazard() const { 1064 return HasLdsBranchVmemWARHazard; 1065 } 1066 1067 bool hasNSAtoVMEMBug() const { 1068 return HasNSAtoVMEMBug; 1069 } 1070 1071 /// Return the maximum number of waves per SIMD for kernels using \p SGPRs 1072 /// SGPRs 1073 unsigned getOccupancyWithNumSGPRs(unsigned SGPRs) const; 1074 1075 /// Return the maximum number of waves per SIMD for kernels using \p VGPRs 1076 /// VGPRs 1077 unsigned getOccupancyWithNumVGPRs(unsigned VGPRs) const; 1078 1079 /// Return occupancy for the given function. Used LDS and a number of 1080 /// registers if provided. 1081 /// Note, occupancy can be affected by the scratch allocation as well, but 1082 /// we do not have enough information to compute it. 1083 unsigned computeOccupancy(const MachineFunction &MF, unsigned LDSSize = 0, 1084 unsigned NumSGPRs = 0, unsigned NumVGPRs = 0) const; 1085 1086 /// \returns true if the flat_scratch register should be initialized with the 1087 /// pointer to the wave's scratch memory rather than a size and offset. 1088 bool flatScratchIsPointer() const { 1089 return getGeneration() >= AMDGPUSubtarget::GFX9; 1090 } 1091 1092 /// \returns true if the machine has merged shaders in which s0-s7 are 1093 /// reserved by the hardware and user SGPRs start at s8 1094 bool hasMergedShaders() const { 1095 return getGeneration() >= GFX9; 1096 } 1097 1098 /// \returns SGPR allocation granularity supported by the subtarget. 1099 unsigned getSGPRAllocGranule() const { 1100 return AMDGPU::IsaInfo::getSGPRAllocGranule(this); 1101 } 1102 1103 /// \returns SGPR encoding granularity supported by the subtarget. 1104 unsigned getSGPREncodingGranule() const { 1105 return AMDGPU::IsaInfo::getSGPREncodingGranule(this); 1106 } 1107 1108 /// \returns Total number of SGPRs supported by the subtarget. 1109 unsigned getTotalNumSGPRs() const { 1110 return AMDGPU::IsaInfo::getTotalNumSGPRs(this); 1111 } 1112 1113 /// \returns Addressable number of SGPRs supported by the subtarget. 1114 unsigned getAddressableNumSGPRs() const { 1115 return AMDGPU::IsaInfo::getAddressableNumSGPRs(this); 1116 } 1117 1118 /// \returns Minimum number of SGPRs that meets the given number of waves per 1119 /// execution unit requirement supported by the subtarget. 1120 unsigned getMinNumSGPRs(unsigned WavesPerEU) const { 1121 return AMDGPU::IsaInfo::getMinNumSGPRs(this, WavesPerEU); 1122 } 1123 1124 /// \returns Maximum number of SGPRs that meets the given number of waves per 1125 /// execution unit requirement supported by the subtarget. 1126 unsigned getMaxNumSGPRs(unsigned WavesPerEU, bool Addressable) const { 1127 return AMDGPU::IsaInfo::getMaxNumSGPRs(this, WavesPerEU, Addressable); 1128 } 1129 1130 /// \returns Reserved number of SGPRs for given function \p MF. 1131 unsigned getReservedNumSGPRs(const MachineFunction &MF) const; 1132 1133 /// \returns Maximum number of SGPRs that meets number of waves per execution 1134 /// unit requirement for function \p MF, or number of SGPRs explicitly 1135 /// requested using "amdgpu-num-sgpr" attribute attached to function \p MF. 1136 /// 1137 /// \returns Value that meets number of waves per execution unit requirement 1138 /// if explicitly requested value cannot be converted to integer, violates 1139 /// subtarget's specifications, or does not meet number of waves per execution 1140 /// unit requirement. 1141 unsigned getMaxNumSGPRs(const MachineFunction &MF) const; 1142 1143 /// \returns VGPR allocation granularity supported by the subtarget. 1144 unsigned getVGPRAllocGranule() const { 1145 return AMDGPU::IsaInfo::getVGPRAllocGranule(this); 1146 } 1147 1148 /// \returns VGPR encoding granularity supported by the subtarget. 1149 unsigned getVGPREncodingGranule() const { 1150 return AMDGPU::IsaInfo::getVGPREncodingGranule(this); 1151 } 1152 1153 /// \returns Total number of VGPRs supported by the subtarget. 1154 unsigned getTotalNumVGPRs() const { 1155 return AMDGPU::IsaInfo::getTotalNumVGPRs(this); 1156 } 1157 1158 /// \returns Addressable number of VGPRs supported by the subtarget. 1159 unsigned getAddressableNumVGPRs() const { 1160 return AMDGPU::IsaInfo::getAddressableNumVGPRs(this); 1161 } 1162 1163 /// \returns Minimum number of VGPRs that meets given number of waves per 1164 /// execution unit requirement supported by the subtarget. 1165 unsigned getMinNumVGPRs(unsigned WavesPerEU) const { 1166 return AMDGPU::IsaInfo::getMinNumVGPRs(this, WavesPerEU); 1167 } 1168 1169 /// \returns Maximum number of VGPRs that meets given number of waves per 1170 /// execution unit requirement supported by the subtarget. 1171 unsigned getMaxNumVGPRs(unsigned WavesPerEU) const { 1172 return AMDGPU::IsaInfo::getMaxNumVGPRs(this, WavesPerEU); 1173 } 1174 1175 /// \returns Maximum number of VGPRs that meets number of waves per execution 1176 /// unit requirement for function \p MF, or number of VGPRs explicitly 1177 /// requested using "amdgpu-num-vgpr" attribute attached to function \p MF. 1178 /// 1179 /// \returns Value that meets number of waves per execution unit requirement 1180 /// if explicitly requested value cannot be converted to integer, violates 1181 /// subtarget's specifications, or does not meet number of waves per execution 1182 /// unit requirement. 1183 unsigned getMaxNumVGPRs(const MachineFunction &MF) const; 1184 1185 void getPostRAMutations( 1186 std::vector<std::unique_ptr<ScheduleDAGMutation>> &Mutations) 1187 const override; 1188 1189 bool isWave32() const { 1190 return WavefrontSize == 32; 1191 } 1192 1193 const TargetRegisterClass *getBoolRC() const { 1194 return getRegisterInfo()->getBoolRC(); 1195 } 1196 1197 /// \returns Maximum number of work groups per compute unit supported by the 1198 /// subtarget and limited by given \p FlatWorkGroupSize. 1199 unsigned getMaxWorkGroupsPerCU(unsigned FlatWorkGroupSize) const override { 1200 return AMDGPU::IsaInfo::getMaxWorkGroupsPerCU(this, FlatWorkGroupSize); 1201 } 1202 1203 /// \returns Minimum flat work group size supported by the subtarget. 1204 unsigned getMinFlatWorkGroupSize() const override { 1205 return AMDGPU::IsaInfo::getMinFlatWorkGroupSize(this); 1206 } 1207 1208 /// \returns Maximum flat work group size supported by the subtarget. 1209 unsigned getMaxFlatWorkGroupSize() const override { 1210 return AMDGPU::IsaInfo::getMaxFlatWorkGroupSize(this); 1211 } 1212 1213 /// \returns Maximum number of waves per execution unit supported by the 1214 /// subtarget and limited by given \p FlatWorkGroupSize. 1215 unsigned getMaxWavesPerEU(unsigned FlatWorkGroupSize) const override { 1216 return AMDGPU::IsaInfo::getMaxWavesPerEU(this, FlatWorkGroupSize); 1217 } 1218 1219 /// \returns Minimum number of waves per execution unit supported by the 1220 /// subtarget. 1221 unsigned getMinWavesPerEU() const override { 1222 return AMDGPU::IsaInfo::getMinWavesPerEU(this); 1223 } 1224 1225 void adjustSchedDependency(SUnit *Src, SUnit *Dst, SDep &Dep) const override; 1226 }; 1227 1228 class R600Subtarget final : public R600GenSubtargetInfo, 1229 public AMDGPUSubtarget { 1230 private: 1231 R600InstrInfo InstrInfo; 1232 R600FrameLowering FrameLowering; 1233 bool FMA; 1234 bool CaymanISA; 1235 bool CFALUBug; 1236 bool HasVertexCache; 1237 bool R600ALUInst; 1238 bool FP64; 1239 short TexVTXClauseSize; 1240 Generation Gen; 1241 R600TargetLowering TLInfo; 1242 InstrItineraryData InstrItins; 1243 SelectionDAGTargetInfo TSInfo; 1244 1245 public: 1246 R600Subtarget(const Triple &TT, StringRef CPU, StringRef FS, 1247 const TargetMachine &TM); 1248 1249 const R600InstrInfo *getInstrInfo() const override { return &InstrInfo; } 1250 1251 const R600FrameLowering *getFrameLowering() const override { 1252 return &FrameLowering; 1253 } 1254 1255 const R600TargetLowering *getTargetLowering() const override { 1256 return &TLInfo; 1257 } 1258 1259 const R600RegisterInfo *getRegisterInfo() const override { 1260 return &InstrInfo.getRegisterInfo(); 1261 } 1262 1263 const InstrItineraryData *getInstrItineraryData() const override { 1264 return &InstrItins; 1265 } 1266 1267 // Nothing implemented, just prevent crashes on use. 1268 const SelectionDAGTargetInfo *getSelectionDAGInfo() const override { 1269 return &TSInfo; 1270 } 1271 1272 void ParseSubtargetFeatures(StringRef CPU, StringRef FS); 1273 1274 Generation getGeneration() const { 1275 return Gen; 1276 } 1277 1278 Align getStackAlignment() const { return Align(4); } 1279 1280 R600Subtarget &initializeSubtargetDependencies(const Triple &TT, 1281 StringRef GPU, StringRef FS); 1282 1283 bool hasBFE() const { 1284 return (getGeneration() >= EVERGREEN); 1285 } 1286 1287 bool hasBFI() const { 1288 return (getGeneration() >= EVERGREEN); 1289 } 1290 1291 bool hasBCNT(unsigned Size) const { 1292 if (Size == 32) 1293 return (getGeneration() >= EVERGREEN); 1294 1295 return false; 1296 } 1297 1298 bool hasBORROW() const { 1299 return (getGeneration() >= EVERGREEN); 1300 } 1301 1302 bool hasCARRY() const { 1303 return (getGeneration() >= EVERGREEN); 1304 } 1305 1306 bool hasCaymanISA() const { 1307 return CaymanISA; 1308 } 1309 1310 bool hasFFBL() const { 1311 return (getGeneration() >= EVERGREEN); 1312 } 1313 1314 bool hasFFBH() const { 1315 return (getGeneration() >= EVERGREEN); 1316 } 1317 1318 bool hasFMA() const { return FMA; } 1319 1320 bool hasCFAluBug() const { return CFALUBug; } 1321 1322 bool hasVertexCache() const { return HasVertexCache; } 1323 1324 short getTexVTXClauseSize() const { return TexVTXClauseSize; } 1325 1326 bool enableMachineScheduler() const override { 1327 return true; 1328 } 1329 1330 bool enableSubRegLiveness() const override { 1331 return true; 1332 } 1333 1334 /// \returns Maximum number of work groups per compute unit supported by the 1335 /// subtarget and limited by given \p FlatWorkGroupSize. 1336 unsigned getMaxWorkGroupsPerCU(unsigned FlatWorkGroupSize) const override { 1337 return AMDGPU::IsaInfo::getMaxWorkGroupsPerCU(this, FlatWorkGroupSize); 1338 } 1339 1340 /// \returns Minimum flat work group size supported by the subtarget. 1341 unsigned getMinFlatWorkGroupSize() const override { 1342 return AMDGPU::IsaInfo::getMinFlatWorkGroupSize(this); 1343 } 1344 1345 /// \returns Maximum flat work group size supported by the subtarget. 1346 unsigned getMaxFlatWorkGroupSize() const override { 1347 return AMDGPU::IsaInfo::getMaxFlatWorkGroupSize(this); 1348 } 1349 1350 /// \returns Maximum number of waves per execution unit supported by the 1351 /// subtarget and limited by given \p FlatWorkGroupSize. 1352 unsigned getMaxWavesPerEU(unsigned FlatWorkGroupSize) const override { 1353 return AMDGPU::IsaInfo::getMaxWavesPerEU(this, FlatWorkGroupSize); 1354 } 1355 1356 /// \returns Minimum number of waves per execution unit supported by the 1357 /// subtarget. 1358 unsigned getMinWavesPerEU() const override { 1359 return AMDGPU::IsaInfo::getMinWavesPerEU(this); 1360 } 1361 }; 1362 1363 } // end namespace llvm 1364 1365 #endif // LLVM_LIB_TARGET_AMDGPU_AMDGPUSUBTARGET_H 1366